語料庫導向之方位短句於固定框架的共現概念統計分析

A Corpus-driven Pattern Analysis in Locative Phrases: A Statistical Comparison of Co-appearing Concepts in Fixed Frames

中文的方位詞組主要可以前飾詞（以、之）與後綴詞（邊、面、頭），結合明確的方向指引（如：前後、上下、左右、裡外等）組合而成。這樣的組成在實際使用上，卻會有避免使用或不存在的組合邏輯，同時這樣的現象亦發生在方位短語構成上。本研究試使用計算統計方法，分析在Sketch Engine中取得的方位名詞組的概念合成模式。在詞彙概念方面，我們使用具知識層級架構的中文同義詞詞林進行將詞彙的概念探索，並計算方位短句裡所包含的知識概念組成模式，最後試從統計方法上尋得詮釋概念與方位詞組組合模式的實證資訊。在本研究之中，我們使用了資訊度量方法中的互斥資訊（Point-wise Mutual Information, PMI）進行統計分析兩個詞組概念間的相關性，並使用多變數互斥資訊（Multivariate Mutual Information, MMI）進行三個概念間的相關分析。本研究的統計結果除了解所選用的語料庫中使用方位名詞的情況外，亦從單一及成對出現的語境概念內容（描述人、物、時空…等在同義詞林中第一階層的名詞），分析各種方位短語使用的前飾/後綴語的搭配方式，以冀期精萃出來的結果，能對方位詞彙的分析上能提供參考的模式。

This paper analyzes synonym groups appearing in fixed frames containing Chinese locative phrases such as [zái noun phrase (yi /zhi ) shàng/xià/etc. bia n/miàn/etc.] by using statistical methods. We collected locative phrases from Sketch Engine using 11 monosyllabic locative words and 5 locative compound-formation patterns, and we aligned these compounds with Chinese Synonym Forest before clustering. Different noun phrases were mapped to their collocating synonym groups to as to enable mutual information comparisons between different combinations. When analyzing concept combinations, we used point-wise mutual information to compare two synonym groups, and adopt multivariate mutual information (MMI)[2] to examine three groups. The results showed that behaviors of using suffixes and prefixes to forming locative nouns in different context (combination of 1 or 2 top level synonym groups), and the statistic results can be used in further analyzing locative nouns in different fields.